A care claim needs more than a comforting chat

A chatbot's claim to provide care needs to explain what happens after its first reassuring reply.

Someone may find a conversation comforting or easier to begin than an appointment. That experience can be useful to them. Clinical assessment needs evidence of its own, with treatment effectiveness evaluated against defined outcomes. A service also needs to account for changing needs.

Research on a developed intervention can supply that evidence under specific conditions. Its findings need to stay attached to the system tested before they can inform a claim about another product.

Identify the intervention in the trial

A March 2025 NEJM AI trial studied Therabot, developed specifically for mental-health treatment. About 200 adults were randomized to access it or join a waitlist. The intervention group reported greater reductions in measured symptoms, with follow-up through eight weeks. The trial abstract

The result applies to that research system. Its training materials form part of the intervention, alongside the participant selection. Safeguards also affect the experience being evaluated, as does the surrounding service.

A waitlist comparison tests a different question from a comparison with established treatment by a qualified professional. Positive ratings of a therapeutic relationship measure another aspect of the experience. Each finding needs the comparison that gives it meaning.

Naming the system limits how far its evidence can travel. A product using the same general category label still needs to establish what its own intervention provides.

Describe the service around the conversation

Dartmouth's account describes expert involvement in Therabot's development. It describes testing for safety and effectiveness. The university calls for further research with mental-health experts closely involved. The account comes from the researchers' institution; independent replication would supply another kind of evidence. Dartmouth's explanation

The interface is one part of the arrangement. Its purpose determines who the intervention is suitable for. The service needs to define what information it can use, then explain how difficulties are handled.

An empathetic style can make a response easy to accept. Assessing clinical appropriateness requires examining its basis against the person's circumstances. A friendly sample conversation leaves that assessment untested.

I'd examine who participated in the evaluation before applying a care claim. The duration also limits the result. The relevant service description needs both beside the outcome measured.

Preserve the clinical limitations

WHO's generative-AI guidance identifies risks from inaccurate information. Biased responses can also cause harm, as can incomplete ones. Its warning about automation bias concerns people overlooking errors or delegating difficult choices because output appears authoritative. WHO's guidance summary

Warm language can carry that same risk. A response may follow an incomplete account of the situation while sounding understanding. Repeating someone's framing leaves its accuracy to be assessed separately.

The American Psychiatric Association's July 2026 position identifies inaccurate or biased information as reasons current general-purpose systems are unreliable for therapeutic care. That is the association's clinical position on those systems. A specialized intervention needs evidence about its own design. The 2026 position statement

A trial of a specialized intervention can report benefit while general-purpose tools remain subject to that warning. The claims concern different systems, each with its own evaluation requirements.

Specify what a handoff completes

Imagine a hypothetical service that flags a need for human attention. It suggests contacting someone, then continues the chat. Sending the suggestion completes a message step; establishing that someone accepted the next step requires evidence of contact.

The service needs to explain whether a person reviews the exchange. It should specify whether users can arrange contact through it, with a stated path when the system reaches its limits.

Capabilities differ between services. Some can connect to human support. The product description needs to state what this service supplies, with evidence about how that connection performs.

Transcript handling also needs a direct explanation. The policy should name who can access it before stating how long it is kept. Permitted uses need specification too, regardless of the conversation's tone.

Give each support task an evaluation

Explaining a service's published information can be assessed against that information. Helping organize questions for an appointment has another purpose. Supporting a defined intervention under oversight requires its own evidence, with a way to detect failures in each task.

Administrative support carries different responsibilities from treatment. Companionship has another purpose again. The people using each service face risks that need to be assessed for that task.

I want a research result with its limits stated beside the benefit. Product descriptions should make those limits understandable before use. A care claim also needs an account of who responds after the conversation has begun.